把 agent 控制判断交给 Jev 小决策模型
作者主张重试、安全、路由、证据是否充分等高频判断不该用大模型生成，而应由 Jev 这类决策模型处理，形成「生成→判断→执行→验证」的架构。
I keep coming back to this after writing about Jev. Maybe we're designing agents around the wrong primitive. Today, one powerful LLM often does everything: reason → choose tool → judge result → decide retry → route model → check safety But most of these aren't generation problems. They're judgment problems. “Should I retry?” “Is this tool call safe?” “Which model should handle this?” “Is the evidence sufficient?” “Did the previous step actually succeed?” Making a frontier LLM answer every tiny control question feels like using a CPU to run every branch of an operating system. This is where models like Jev become interesting. A generative model can handle the expensive reasoning. A small decision model can handle high-frequency, bounded judgments. Code handles state, permissions and side effects. Humans handle the ambiguous/high-risk cases. So the agent architecture becomes: Generate → Judge → Act → Verify → Generate rather than: LLM → LLM → LLM → LLM → ... And I think there's a huge optimization opportunity here: Don't make the model smarter at everything. Make the system smarter about which model gets asked what. That could mean lower latency, lower cost, better observability, and potentially much more predictable agent behavior. The interesting future question isn't: “How powerful should our agent model be?” It's: “Which decisions actually deserve an expensive model?” That feels like a fundamental agent-harness problem. https://www.harness.institute/research/jev-decision-model-in-the-agent-harness
这条是引用着另一条写的：「文章：Jev 是 AI Agent 缺失的决策层」